AdaptiveFrequencyNet: a long-term time series forecasting approach based on learnable frequency decomposition and...

Area of Science:

• Machine Learning
• Time Series Analysis
• Deep Learning

Background:

• Long-term time series forecasting requires modeling variations across multiple time scales and temporal dependencies.
• Existing frequency decomposition methods have limitations in multi-scale modeling, feature fusion, and cross-scale interaction.
• Fixed frequency splits can negatively impact long-horizon forecasting consistency.

Purpose of the Study:

• To introduce AdaptiveFrequencyNet (AFNet), a novel architecture for enhanced long-term time series forecasting.
• To address limitations of existing methods by incorporating learnable multi-scale filter banks and cross-scale attention fusion.
• To evaluate AFNet's performance against state-of-the-art models on standard forecasting benchmarks.

Main Methods:

• AFNet utilizes learnable time-domain filter banks via depthwise convolutions and branch-wise softmax mixing.
• It employs Adaptive Multi-Scale Feature Fusion (AMSFF) combining self-attention, multi-scale convolutions, and gating.
• Cross-Scale Attention Fusion (CSAF) facilitates progressive mixing of information across different scales using cross-attention.

Main Results:

• AFNet achieves competitive test errors on six standard benchmarks with a moderate parameter budget.
• Performance is dataset- and horizon-dependent, with AFNet showing strong results comparable to iTransformer and Pathformer on specific datasets.
• Ablation studies confirm the significant contribution of each module within the AFNet architecture.

Conclusions:

• AFNet offers an effective approach to multi-scale time series modeling, outperforming or matching strong baselines.
• The architecture's adaptive nature and cross-scale attention mechanism are key to its forecasting capabilities.
• Further research can explore the application of AFNet to diverse time series forecasting challenges.